Skip to content

Repository files navigation

Pastewatch

Stable Version License: MIT CI ANCC

Detects and obfuscates sensitive data before it reaches AI systems — clipboard monitoring, CLI scanner, MCP server, API proxy, shell guard hooks, and VS Code extension.

It operates before paste, not after submission.

If sensitive data never enters the prompt, the incident does not exist.


Core Principle

Principiis obsta - resist the beginnings.

Pastewatch intervenes at the earliest irreversible boundary: the moment data leaves the user's control.

Once pasted into an AI system, data cannot be reliably recalled, audited, or constrained.

Pastewatch refuses that transition.


Why This Matters

Every AI agent sends your file contents, command outputs, and tool results to a cloud API. If those contain secrets, the secrets leave your machine — silently, irreversibly, and into infrastructure you don't control.

Pastewatch prevents supported secret-leakage paths structurally without breaking agent workflows:

  What the agent does                What actually happens
  ──────────────────                 ──────────────────────
  Read a file with secrets     →    MCP returns placeholders, secrets stay in RAM
  Run a bash command with DSN  →    Guard blocks before execution
  Send tool results to API     →    Proxy redacts secrets from the request body
  Write code with placeholders →    MCP resolves originals locally on write-back

The agent works normally. It reads files, runs commands, and writes code. Through MCP, the agent sees reversible placeholders; through the proxy, intrinsically identifiable or operator-authorized secrets are replaced before supported traffic reaches the cloud.

No ML and no probabilistic mutation. Authorization comes from secret evidence, not severity.

Why Pastewatch

  • Before-paste boundary — authorized secret matches are rewritten before supported traffic leaves. Nightfall, Prisma, Check Point all intercept downstream. Pastewatch prevents upstream
  • MCP server for AI agents — no other tool provides redacted read/write at the tool level. Authorized matches become reversible placeholders while the secret map stays local
  • Bash guard with deep parsing — pipes, subshells, redirects, database CLIs, infra tools. Every shell command the agent runs is scanned before execution
  • API proxy — catches Anthropic-shaped traffic that bypasses hooks, including from subagents. Last line of defense before the network boundary (refuses unrecognized upstream shapes rather than forward them unscanned)
  • Canary honeypots — "prove it works" not "trust it works." Plant format-valid fake secrets and verify they're caught
  • Local-only, deterministic, no ML — no cloud dependency, no probabilistic scoring, no telemetry. Runs offline, gives the same answer every time
  • One commandpastewatch-cli launch claude and every layer is active. No manual setup, no env vars, no second terminal

What Pastewatch Does

Pastewatch started as a clipboard monitor — scan before paste, replace secrets with placeholders. It evolved into a full secret protection stack for AI agent workflows:

Layer What it does How it works
Clipboard monitor Scans before paste macOS menubar app, replaces secrets in clipboard
CLI scanner Scans files, directories, git diffs pastewatch-cli scan --dir .
Startup sweep Warns about pre-existing shell config credentials pastewatch-cli launch scans common startup files once per changed finding summary
MCP server Redacted read/write for AI agents Agent sees placeholders, originals stay in RAM
Shell guard Blocks secrets in commands and file access Pre-execution hook for Claude Code, Cline, Cursor, Windsurf, Continue, Amazon Q
API proxy Redacts secrets from supported outbound API traffic Scans Anthropic-shaped requests and refuses unsupported body shapes
VS Code extension Real-time detection in the editor Highlights secrets as you type

All layers share the same detection engine — 30+ pattern types, deterministic regex, no ML. Every layer operates locally. Nothing phones home.

Zero false-positive mutation

Pastewatch rewrites your data only when it is certain the value is a secret. This is a hard rule, not a tuning knob:

  • Mutated: intrinsically identifiable secrets such as sourced provider tokens, complete private keys, validated JWTs and cards; exact values supplied by a trusted local source; patterns you approve with a custom rule; and email/host values covered by an obfuscate entry.
  • Opt-in ambiguity: intrinsic detectors are enabled by default. Email and host mutation requires an exact or domain-scoped obfuscate entry. Unconfigured email/host values are not rewritten or nagged; the proxy records only privacy-safe domain coverage so the operator can decide whether to opt in. Other ambiguous detectors remain advisory-only when explicitly enabled.
  • --severity controls how much it nags, never what it rewrites. Lowering severity surfaces more advisories; it never widens the set of values that get mutated.

The result: false negatives are preferred over false positives, and mutation false positives are driven to ~zero by construction. Pastewatch never breaks a working agent response to redact something it only might be.


What Pastewatch is NOT

  • Not a DLP system — no policies, no enforcement workflows, no admin console
  • Not a compliance product — it does not certify, audit, or generate reports for regulators
  • Not an AI classifier — deterministic pattern matching only, no probabilistic scoring
  • Not a policy engine — it does not decide what you're allowed to do, it prevents structural leaks

Pastewatch does not:

  • phone home or collect telemetry
  • require cloud connectivity
  • guess, infer, or act when uncertain
  • store clipboard history or file contents
  • make decisions — it presents evidence and lets you decide

How Pastewatch Works

Pastewatch scans text for sensitive patterns and replaces them with non-sensitive placeholders. The same engine powers all six layers:

  1. Detection — regex-based pattern matching across 30+ secret types (API keys, DSNs, tokens, credentials, PII)
  2. Authorization — evidence partitions matches into mutation or advisory-only outcomes
  3. Obfuscation — authorized values are replaced with typed placeholders (<AWS_KEY_1>)
  4. Resolution — only the MCP server stores originals in local RAM and restores them on write-back

The clipboard monitor scans before paste. The CLI scans files on demand. The MCP server scans on read and resolves on write. The guard scans commands before execution. The proxy scans API requests before they leave the network. Each layer catches what the others miss.


Quick Start

30 seconds from zero to protected AI agent session:

# 1. Install
brew install ppiankov/tap/pastewatch

# 2. Set up hooks and MCP server for your agent
pastewatch-cli setup claude-code

# 3. Run through the proxy — one command, fully protected
pastewatch-cli launch claude

The launch command starts the proxy, waits for it to be ready, sets ANTHROPIC_BASE_URL, and runs Claude Code. When the agent exits, the proxy stops. The proxy scans Anthropic-shaped API requests and redacts supported secrets before they leave your machine. Protect other agents with configured hooks, MCP tools, and agent instructions where available.

Important: The setup step injects credential handling rules into your agent's CLAUDE.md. Without these rules, agents may echo passwords in shell output or store plaintext credentials in memory files — formats that bypass regex detection. The rules ensure agents use detectable keywords (password=, secret=) and never store raw values. See docs/CLAUDE-SNIPPET.md for the full snippet.

For persistent setup, add a shell alias:

# .zshrc / .bashrc
alias claude='pastewatch-cli launch claude'

Installation

From Release (Recommended)

  1. Download the latest .dmg from Releases
  2. Open the DMG and drag Pastewatch.app to Applications
  3. Launch Pastewatch from Applications
  4. Grant notification permissions when prompted

CLI via Homebrew

brew install ppiankov/tap/pastewatch
pastewatch-cli doctor    # verify installation

CLI Manual Install (No Homebrew)

For environments where Homebrew is not available (CI runners, restricted workstations):

# macOS (universal binary — Apple Silicon + Intel)
curl -L -o pastewatch-cli https://github.com/ppiankov/pastewatch/releases/latest/download/pastewatch-cli

# Linux x86_64
curl -L -o pastewatch-cli https://github.com/ppiankov/pastewatch/releases/latest/download/pastewatch-cli-linux-amd64

# Linux arm64
curl -L -o pastewatch-cli https://github.com/ppiankov/pastewatch/releases/latest/download/pastewatch-cli-linux-arm64

chmod +x pastewatch-cli
sudo mv pastewatch-cli /usr/local/bin/
pastewatch-cli doctor

Or build from source (requires Swift 5.9+):

git clone https://github.com/ppiankov/pastewatch.git
cd pastewatch
swift build -c release
sudo cp .build/release/PastewatchCLI /usr/local/bin/pastewatch-cli

From Source (GUI)

git clone https://github.com/ppiankov/pastewatch.git
cd pastewatch
swift build -c release
./.build/release/pastewatch

Detection Scope

Pastewatch detects only deterministic, high-confidence patterns:

Type Examples
Email user@company.com
Phone +60123456789, (555) 123-4567
IP Address 192.168.1.100
AWS Keys AKIAIOSFODNN7EXAMPLE
API Keys sk_test_..., ghp_...
UUIDs 550e8400-e29b-41d4-a716-446655440000
JWT Tokens eyJhbGciOiJIUzI1NiIs...
DB Connections postgres://..., clickhouse://...
SSH Keys -----BEGIN RSA PRIVATE KEY-----
Credit Cards 4111111111111111 (Luhn validated)
File Paths /etc/nginx/nginx.conf, /home/deploy/.ssh/id_rsa
Hostnames db-primary.internal.corp.net
Credentials password=..., secret: ..., api_key=...
Slack Webhooks https://hooks.slack.com/services/...
Discord Webhooks https://discord.com/api/webhooks/...
Azure Connections DefaultEndpointsProtocol=https;AccountName=...;AccountKey=...
GCP Service Accounts {"type": "service_account", ...}
OpenAI Keys sk-proj-..., sk-svcacct-...
Anthropic Keys sk-ant-api03-..., sk-ant-admin01-...
Hugging Face Tokens hf_...
Groq Keys gsk_...
npm Tokens npm_...
PyPI Tokens pypi-...
RubyGems Tokens rubygems_...
GitLab Tokens glpat-...
Telegram Bot Tokens 123456789:AA...
SendGrid Keys SG....
Shopify Tokens shpat_..., shpca_...
DigitalOcean Tokens dop_v1_..., doo_v1_...
Perplexity Keys pplx-...
JDBC URLs jdbc:oracle:thin:@..., jdbc:db2://..., jdbc:postgresql://...
XML Credentials <password>, <secret_access_key>, etc. in XML configs
XML Usernames <user>, <quota_key> in XML configs
XML Hostnames <host>, <hostname>, <interserver_http_host> in XML configs
High Entropy Strings Opt-in Shannon entropy detection (4.0 bits/char threshold)

Each type has a severity level (critical, high, medium, low) used in SARIF, JSON, and markdown output.

No ML. No probabilistic scoring. No confidence levels.

If detection is ambiguous, Pastewatch does nothing.


Obfuscation Model

Detected values are replaced with typed, numbered placeholders:

AKIAIOSFODNN7EXAMPLE  →  <AWS_KEY_1>
AIza...               →  <GOOGLE_API_KEY_1>
github_pat_...        →  <GITHUB_TOKEN_1>

How placeholders work depends on the layer:

Layer Placeholder lifetime Recovery
Clipboard Discarded after paste None — one-way
CLI scan Output only None — source files are never modified
MCP server Stored in RAM for the session Write-back resolves originals locally
API proxy Replaced in-flight None — redacted before it leaves

The MCP server is the only layer that maintains a mapping, because it must restore placeholders locally when the agent writes a file. Clipboard and proxy replacement is one-way; responses are scanned independently rather than deobfuscated. The MCP mapping lives in process memory and is lost when the session ends.


User Experience

  • Clipboard/GUI — silent by default. When obfuscation occurs, a minimal macOS notification: Pastewatch: Obfuscated: AWS Key (1), Google API Key (1)
  • CLI — findings printed to stdout, exit code 6 if secrets found
  • Startup sweep — one stderr warning per changed shell config finding summary during launch; disable with --no-startup-sweep (details)
  • MCP — transparent to the agent. It reads placeholders and writes them back. No user interaction needed
  • Guard hook — blocks with a clear message: BLOCKED: file contains secrets. Use pastewatch_read_file instead
  • Proxy — redacts silently. When secrets are caught, injects a [PASTEWATCH] alert into the agent's response so it can warn the user

No previews. No animations. No confirmations. Silence is success.


CLI Mode

pastewatch-cli provides scanning, guarding, and proxy subcommands for use without the GUI:

pastewatch-cli scan --dir .              # scan a directory
pastewatch-cli launch claude             # proxy + agent in one step
pastewatch-cli guard "cat .env"          # block secret-leaking commands
pastewatch-cli mcp                        # redacted read/write MCP server

Full command reference: docs/cli-reference.md — every subcommand (scan, proxy, launch, mcp, guard, fix, inventory, report, canary, watch, dashboard, config, and CI integration) with flags and examples.

Agent Integration

Agent Safety Matrix

The API proxy (Layer 0) redacts supported Anthropic-shaped traffic (/v1/messages and /v1/messages/batches) and refuses unrecognized upstream body shapes (HTTP 415). pastewatch-cli launch wires only Claude Code to this proxy. Other clients need hooks or MCP unless the operator separately configures an Anthropic-compatible client to use pastewatch-cli proxy; OpenAI/Gemini-shaped bodies are not redacted.

Agent Protection Hooks MCP Setup
Claude Code Structural PreToolUse (exit 2) Automatic pastewatch-cli setup claude-code
Cline Structural PreToolUse (JSON cancel) Automatic pastewatch-cli setup cline
Roo Code Structural PreToolUse (JSON cancel) Automatic pastewatch-cli setup roo-code
Cursor Structural preToolUse (exit 2) Automatic pastewatch-cli setup cursor
Windsurf Structural pre_read/write/run (exit 2) Automatic pastewatch-cli setup windsurf
Continue Structural PreToolUse (exit 2) Automatic pastewatch-cli setup continue
Amazon Q Structural preToolUse (exit 2) Automatic pastewatch-cli setup amazon-q
Copilot Structural preToolUse (.github/hooks/) Automatic pastewatch-cli setup copilot
Codex CLI Structural after hook trust PreToolUse (exit 2) Manual TOML pastewatch-cli setup codex, then review with /hooks
Qwen Code Structural PreToolUse (exit 2) Automatic pastewatch-cli setup qwen-code
Goose MCP only No hooks Manual YAML pastewatch-cli setup goose
Kilo Code MCP only No hooks (declined) Automatic pastewatch-cli setup kilo-code
Gemini MCP only No hooks Automatic pastewatch-cli setup gemini
Antigravity (agy) Proxy not applicable + MCP only (voluntary tools; no Structural read blocking) No - handler registration fails Yes Not yet supported by pastewatch-cli setup; edit ~/.gemini/config/mcp_config.json manually
OpenCode MCP only No hooks Yes Manual
Aider No automatic local layer No MCP yet Unavailable pastewatch-cli setup aider explains the limitation
Jules Cloud only No local config Cloud UI N/A (use proxy on local side)

Structural = hooks block native file access before secrets can be read. The agent cannot bypass the check. MCP only = redaction applies only when the agent uses the configured Pastewatch MCP tools. Native file or network access is not intercepted.

Install

Install via Homebrew:

brew install ppiankov/tap/pastewatch

Or download the binary:

curl -LO https://github.com/ppiankov/pastewatch/releases/latest/download/pastewatch-cli
chmod +x pastewatch-cli
sudo mv pastewatch-cli /usr/local/bin/

For AI coding agents: Use MCP redacted read/write to prevent secret leakage - see docs/agent-safety.md for setup.

For CI/CD: Use the CLI scan command or GitHub Action.

Agents: read docs/SKILL.md for commands, flags, config files, detection types, and exit codes.


Configuration

Config files

File Location Purpose Created By
.pastewatch.json Project root Project-level config (rules, allowlists, hosts) pastewatch-cli init
~/.config/pastewatch/config.json Home User-level defaults Manual / GUI app
.pastewatch-allow Project root Value allowlist (one per line, # comments) pastewatch-cli init
.pastewatchignore Project root Path exclusion patterns (glob, like .gitignore) Manual
.pastewatch-baseline.json Project root Known findings baseline pastewatch-cli baseline create

Resolution cascade: CWD .pastewatch.json > ~/.config/pastewatch/config.json > built-in defaults.

.pastewatch.json schema

{
  "enabled": true,
  "enabledTypes": ["AWS Key", "SSH Private Key", "JWT Token", "Vault Token"],
  "showNotifications": true,
  "soundEnabled": false,
  "allowedValues": ["test@example.com"],
  "allowedPatterns": ["sk_test_.*", "EXAMPLE_.*"],
  "customRules": [
    {"name": "Internal ID", "pattern": "MYCO-[0-9]{6}", "severity": "medium"}
  ],
  "safeHosts": [".internal.company.com"],
  "sensitiveHosts": [".local", "secrets.vault.internal.net"],
  "sensitiveIPPrefixes": ["172.16.", "10."],
  "mcpMinSeverity": "high",
  "operatorRedactionNotices": false,
  "obfuscate": [
    {"type": "email", "pattern": "operator@example.com"},
    {"type": "email", "pattern": "@example.com"},
    {"type": "host", "pattern": "gateway.example.com"},
    {"type": "host", "pattern": ".example.com"}
  ],
  "placeholderPrefix": "REDACTED_PLACEHOLDER_"
}
Field Type Description
enabled bool Enable/disable scanning globally
enabledTypes string[] Detection types to activate (default: intrinsic detectors only)
showNotifications bool System notifications on GUI obfuscation
soundEnabled bool Sound on GUI obfuscation
allowedValues string[] Exact values to suppress (merged with .pastewatch-allow)
allowedPatterns string[] Regex patterns for value suppression (wrapped in ^(...)$)
customRules object[] Additional regex patterns with name, pattern, optional severity
safeHosts string[] Hostnames excluded from detection (leading dot = suffix match)
sensitiveHosts string[] Hostnames always detected (overrides safe hosts, catches 2-segment hosts like .local)
sensitiveIPPrefixes string[] IP prefixes always detected (overrides built-in exclude list, e.g., 172.16.)
mcpMinSeverity string Default severity threshold for MCP redacted reads (default: high)
operatorRedactionNotices bool Emit a metadata-only operator notice for every proxy/MCP mutation, including under --quiet (default: false)
obfuscate object[] Exact or domain-scoped email/host mutation rules. Email patterns are an address or @domain; host patterns are a hostname or .domain. Invalid entries fail configuration validation.
placeholderPrefix string? Custom prefix for MCP placeholders (e.g., REDACTED_ produces REDACTED_001). Default: null (uses __PW_TYPE_N__ format)

GUI settings can also be changed via the menubar dropdown.


Threat Model

Pastewatch assumes:

  • Users will paste sensitive data
  • AI systems are not trusted with raw secrets
  • Prevention is cheaper than remediation

Pastewatch does not attempt to secure downstream systems. It prevents supported, positively identified secrets from entering configured agent transports; ambiguous classes stay unchanged unless the operator authorizes them.


Design Constraints

  • Local-only operation
  • Deterministic behavior
  • Minimal UI surface
  • No background analytics
  • No user accounts
  • No configuration required for safe defaults

If a feature increases complexity without reducing risk, it is rejected.


Platform Support

Platform Component Status
macOS (Apple Silicon) GUI + CLI Supported
macOS (Intel x86_64) CLI only Supported (universal binary)
Linux x86_64 CLI only Supported
Linux arm64 CLI only Supported
Windows Not supported natively — run the Linux binary under WSL2

The GUI (clipboard monitoring) is macOS-only. The CLI runs on macOS and Linux via Homebrew or direct download.

Windows: there is no native Windows build. Run the CLI inside WSL2 using the Linux amd64 binary (see Installation → Linux x86_64). If you need native Windows secret-redaction on the API wire, that is the domain of the gateway layer, not this tool.

Linux proxy note: On Linux, the proxy uses curl (Process) for upstream HTTP requests instead of Swift's URLSession. This works around a known FoundationNetworking bug on arm64 where dataTask completion handlers silently fail. Requires curl on PATH (present on all standard Linux distributions).


Documentation


License

MIT License.

Use it. Fork it. Modify it.

Do not pretend it guarantees compliance or safety.


Project Status

Status: Stable, feature-complete · v0.36.1 · Accepting compatibility and bug fixes only

Pastewatch is stable and in maintenance mode. See docs/status.md for the full feature-milestone breakdown.

About

Detects and obfuscates sensitive data before it reaches AI systems — clipboard, CLI, and MCP server

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages